Cannot load the last model (model_5) in the Milestone project 2 Skimlit

[BaoLocPham]

I'm stuck with two error for 2 days and still cannot load the model.

1. I train model by my self but when i load it, there is an error:
   this is the warning when i save my model
   
   this is the error:
   

2. I loaded @mrdbourke pretrain model but there is an erorr:
   

What i haved tried:

• I changed the tensorflow version to 2.5 and 2.4 but it doesn't work.
• I changed the code for model 5 a little bit, but it doesn't work :
  
• Although i run the template notebook but it doesn't work.

[mrdbourke]

Hmm... I haven't seen either of those before.

Thank you for uploading what you've tried so far.

Has anything else worked? Or not worked?

I'll dig into the template notebook tomorrow and see what's going on - I'll post my findings here.

Update:

Looks like it's a TensorFlow 2.5+ issue:

WARNING:tensorflow:SavedModel saved prior to TF 2.5 detected when loading Keras model. Please ensure that you are saving the model with model.save() or tf.keras.models.save_model(), *NOT* tf.saved_model.save(). To confirm, there should be a file named "keras_metadata.pb" in the SavedModel directory.

Will get deeper into this tomorrow. I think the model on GCP (the one downloaded into the notebook doesn't contain the keras_metadata.pb file TensorFlow is after.

[mrdbourke]

Hmm, still working on this.

I'm getting the same errors you're getting with the RaggedTensors, I've not seen these before and searching online doesn't offer much.

May have to post on the forums - https://discuss.tensorflow.org/

[mrdbourke]

I've found a fix!

It looks like the masking was an issue indeed.

This error:

The mask that was passed in was tf.RaggedTensor(values=Tensor("Placeholder_2:0", shape=(None,), dtype=bool), row_splits=Tensor("Placeholder_3:0", shape=(None,), dtype=int64)) and cannot be applied to RaggedTensor inputs. Please make sure that there is no mask passed in by upstream layers.

Turns out the masking happened in the char_embed creation code:

# Create char embedding layer (OLD)
char_embed = layers.Embedding(input_dim=NUM_CHAR_TOKENS, # number of different characters
                              output_dim=25, # embedding dimension of each character (same as Figure 1 in https://arxiv.org/pdf/1612.05251.pdf)
                              mask_zero=True, # <-- mask creation 
                              name="char_embed")

To remove the mask, I set mask_zero=False:

# Create char embedding layer (NEW) 
char_embed = layers.Embedding(input_dim=NUM_CHAR_TOKENS, # number of different characters
                              output_dim=25, # embedding dimension of each character (same as Figure 1 in https://arxiv.org/pdf/1612.05251.pdf)
                              mask_zero=False, # don't use masks (this messes up model_5 if set to True)
                              name="char_embed")

This revolves the loading and saving of model_5.

Also, I've updated the model stored in Google Storage (same URL download as in the course) to include the new keras_metadata.pb file included in saved TensorFlow models from TensorFlow 2.5+.

With keras_metadata.pb (this should be created as long as you use model.save()), the theory is that you no longer need the custom_objects parameter in load_model():

# Old
loaded_model = tf.keras.models.load_model("skimlit_gs_model/skimlit_tribrid_model",
    custom_objects={"TextVectorization": TextVectorization,
                                   "KerasLayer": hub.KerasLayer}) 

# New (with keras_metadata.pb, no custom_objects parameter)
loaded_model = tf.keras.models.load_model("skimlit_gs_model/skimlit_tribrid_model")

I've updated the code in notebook 09 to reflect all of these changes - it should now run top to bottom.

[BaoLocPham]

Thank you @mrdbourke , after i changed the code as your solution, it worked.

[rkrish999]

I also had the same problem and tried the same fixes as BoaLocPham. Thanks for pposting the fix here @mrdbourke!